Skip to Content
DocumentationGetting startedChoose hardware for Companion OS

Choose hardware for Companion OS

Companion OS runs on an ordinary 64-bit PC. This page helps you decide whether a spare machine, a mini PC, or a small server is enough, and what to expect from it.

Current Companion OS testing uses virtual machines rather than physical hardware. Wi-Fi, Bluetooth, sound, suspend, and GPU behavior on physical machines are still under evaluation. Treat the sizes below as planning numbers, and report what you find on your machine (see Troubleshooting Companion OS).

Pick the image for your machine

Your machineImageNotes
Intel or AMD PC (x86-64)companion-os-<version>-amd64The reference Ubuntu 26.04.1 LTS image (GNOME desktop). Includes Companion Capture.
Intel or AMD PC (x86-64, KDE Plasma)companion-os-plasma-<YYYY.MM.DD>-x86_64Alternative CachyOS-based image with KDE Plasma 6 and the Calamares online installer.
arm64 PC, server, or virtual machinecompanion-os-<version>-arm64Beta Ubuntu 26.04.1 LTS image. Does not include Companion Capture yet because no Linux arm64 build exists.
Mac with Apple siliconnoneA Mac cannot boot from the USB installer. Run the arm64 image in a virtual machine instead.

The machine must start in UEFI mode and have an internal disk. The USB stick serves only as the installer.

What the machine needs

MinimumComfortable
Processor64-bit, 2 cores4 cores or more
Memory4 GB8 GB or more
Disk32 GB (the one-click installer refuses disks under 25 GB)64 GB or more, an SSD
NetworkEthernet or Wi-Fi, internet on the first startEthernet
Display1280 by 800Any. You need a screen for the install.

A test machine with 8 GB of memory and a 64 GB disk finished the first start with about 19 GB of disk space and 3 GB of memory in use, including Companion Hub and two small models.

Running without a monitor after installation? Companion Hub is a web application, so once you install the OS and connect the machine to your network, you can manage it from another device. The OS installation itself requires a screen and a keyboard. For machines without a display, Headless Wi-Fi setup covers the setup hotspot on pre-installed appliances.

Which AI models you get

The first-time setup downloads the default appliance-core model pack that fits your hardware:

ModelSizeWhat it is forWhich machines
nomic-embed-text274 MBEmbeddings (search and memory)All machines
gemma3:1b806 MBA small chat modelAll machines
qwen3.6:27b17.8 GBA large chat modelAMD machines with unified memory (Ryzen AI Max / Strix Halo class) and at least 30 GB of RAM

Every other machine downloads the first two models (about 1.1 GB in total). The small chat model runs on a CPU for testing Companion Hub, though larger tasks benefit from a bigger model.

To use larger models, pull them with Ollama (ollama pull <model>). Make sure your system RAM or GPU VRAM can hold the model weights. See Inference and AI for how Hub routes requests to local models.

Graphics cards

GPUWhat to do
AMD Ryzen AI Max (Strix Halo)On the amd64 Ubuntu image, run sudo ci-os-install-rocm to install AMD’s ROCm compute stack (gfx1151) from the Ubuntu archive, then reboot.
NVIDIACompanion OS does not pre-install proprietary NVIDIA drivers. To install one on the Ubuntu image, run sudo ubuntu-drivers install and restart. Ollama uses a working NVIDIA driver automatically.
Intel, other AMD, no discrete GPUModels run on the CPU.

Plan your disk

The operating system, Docker images for Hub, and the two small models use about 19 GB. Hub apps and additional models add to that footprint. A 64 GB disk is a practical starting point, and 256 GB or more is recommended if you plan to store large models or media libraries.

The installer erases the entire target disk. Back up first, and choose a disk that holds nothing you want to keep.

Next

Last updated on